def dense_layer_3d_proj(input_tensor,
hidden_size,
head_size,
initializer,
activation,
name=None):
"""A dense layer with 3D kernel for projection.
Args:
input_tensor: float Tensor of shape [batch,from_seq_length,
num_attention_heads, size_per_head].
hidden_size: The size of hidden layer.
num_attention_heads: The size of output dimension.
head_size: The size of head.
initializer: Kernel initializer.
activation: Actication function.
name: The name scope of this layer.
Returns:
float logits Tensor.
"""
input_shape = albert_utils_official.get_shape_list(input_tensor)
num_attention_heads= input_shape[2]
with tf.variable_scope(name):
w = tf.get_variable(
name="kernel",
shape=[num_attention_heads * head_size, hidden_size],
initializer=initializer)
w = tf.reshape(w, [num_attention_heads, head_size, hidden_size])
b = tf.get_variable(
name="bias", shape=[hidden_size], initializer=tf.zeros_initializer)
ret = tf.einsum("BFND,NDH->BFH", input_tensor, w)
ret += b
if activation is not None:
return activation(ret)
else:
return ret
def dense_layer_2d(input_tensor,
output_size,
initializer,
activation,
num_attention_heads=1,
name=None):
"""A dense layer with 2D kernel.
Args:
input_tensor: Float tensor with rank 3.
output_size: The size of output dimension.
initializer: Kernel initializer.
activation: Activation function.
num_attention_heads: number of attention head in attention layer.
name: The name scope of this layer.
Returns:
float logits Tensor.
"""
del num_attention_heads # unused
input_shape = albert_utils_official.get_shape_list(input_tensor)
hidden_size = input_shape[2]
with tf.variable_scope(name):
w = tf.get_variable(
name="kernel",
shape=[hidden_size, output_size],
initializer=initializer)
b = tf.get_variable(
name="bias", shape=[output_size], initializer=tf.zeros_initializer)
ret = tf.einsum("BFH,HO->BFO", input_tensor, w)
ret += b
if activation is not None:
return activation(ret)
else:
return ret
def dot_product_attention(q, k, v, bias, dropout_rate=0.0):
"""Dot-product attention.
Args:
q: Tensor with shape [..., length_q, depth_k].
k: Tensor with shape [..., length_kv, depth_k]. Leading dimensions must
match with q.
v: Tensor with shape [..., length_kv, depth_v] Leading dimensions must
match with q.
bias: bias Tensor (see attention_bias())
dropout_rate: a float.
Returns:
Tensor with shape [..., length_q, depth_v].
"""
logits = tf.matmul(q, k, transpose_b=True) # [..., length_q, length_kv]
logits = tf.multiply(
logits,
1.0 / math.sqrt(float(albert_utils_official.get_shape_list(q)[-1])))
if bias is not None:
# `attention_mask` = [B, T]
from_shape = albert_utils_official.get_shape_list(q)
if len(from_shape) == 4:
broadcast_ones = tf.ones([from_shape[0], 1, from_shape[2], 1],
tf.float32)
elif len(from_shape) == 5:
# from_shape = [B, N, Block_num, block_size, depth]#
broadcast_ones = tf.ones(
[from_shape[0], 1, from_shape[2], from_shape[3], 1],
tf.float32)
bias = tf.matmul(broadcast_ones,
tf.cast(bias, tf.float32),
transpose_b=True)
# Since attention_mask is 1.0 for positions we want to attend and 0.0 for
# masked positions, this operation will create a tensor which is 0.0 for
# positions we want to attend and -10000.0 for masked positions.
adder = (1.0 - bias) * -10000.0
# Since we are adding it to the raw scores before the softmax, this is
# effectively the same as removing these entirely.
logits += adder
else:
adder = 0.0
attention_probs = tf.nn.softmax(logits, name="attention_probs")
attention_probs = dropout(attention_probs, dropout_rate)
return tf.matmul(attention_probs, v)
def build_encoder(self, input_ids, input_mask, hidden_dropout_prob,
attention_probs_dropout_prob, **kargs):
reuse = kargs["reuse"]
input_shape = albert_utils_official.get_shape_list(
input_ids, expected_rank=[2, 3])
batch_size = input_shape[0]
seq_length = input_shape[1]
if input_mask is None:
input_mask = tf.ones(shape=[batch_size, seq_length],
dtype=tf.int32)
with tf.variable_scope(self.config.get("scope", "bert"), reuse=reuse):
with tf.variable_scope("encoder"):
# This converts a 2D mask of shape [batch_size, seq_length] to a 3D
# mask of shape [batch_size, seq_length, seq_length] which is used
# for the attention scores.
# Run the stacked transformer.
# `sequence_output` shape = [batch_size, seq_length, hidden_size].
if kargs.get('attention_type',
'efficient_attention') == 'normal_attention':
tf.logging.info("****** normal attention *******")
transformer_model = albert_modules_official.transformer_model
else:
tf.logging.info("****** normal attention *******")
transformer_model = albert_modules_official.transformer_model
[
self.all_encoder_layers, self.all_attention_scores
] = albert_modules_official.transformer_model(
input_tensor=self.embedding_output,
attention_mask=input_mask,
hidden_size=self.config.hidden_size,
num_hidden_layers=self.config.num_hidden_layers,
num_hidden_groups=self.config.num_hidden_groups,
num_attention_heads=self.config.num_attention_heads,
intermediate_size=self.config.intermediate_size,
inner_group_num=self.config.inner_group_num,
intermediate_act_fn=albert_modules_official.get_activation(
self.config.hidden_act),
hidden_dropout_prob=self.config.hidden_dropout_prob,
attention_probs_dropout_prob=self.config.
attention_probs_dropout_prob,
initializer_range=self.config.initializer_range,
do_return_all_layers=True)
def embedding_lookup(input_ids, vocab_size, embedding_size=128, initializer_range=0.02, word_embedding_name="word_embeddings", use_one_hot_embeddings=False): """Looks up words embeddings for id tensor. Args: input_ids: int32 Tensor of shape [batch_size, seq_length] containing word ids. vocab_size: int. Size of the embedding vocabulary. embedding_size: int. Width of the word embeddings. initializer_range: float. Embedding initialization range. word_embedding_name: string. Name of the embedding table. use_one_hot_embeddings: bool. If True, use one-hot method for word embeddings. If False, use `tf.nn.embedding_lookup()`. Returns: float Tensor of shape [batch_size, seq_length, embedding_size]. """ # This function assumes that the input is of shape [batch_size, seq_length, # num_inputs]. # # If the input is a 2D tensor of shape [batch_size, seq_length], we # reshape to [batch_size, seq_length, 1]. if input_ids.shape.ndims == 2: input_ids = tf.expand_dims(input_ids, axis=[-1]) embedding_table = tf.get_variable( name=word_embedding_name, shape=[vocab_size, embedding_size], initializer=create_initializer(initializer_range)) if use_one_hot_embeddings: flat_input_ids = tf.reshape(input_ids, [-1]) one_hot_input_ids = tf.one_hot(flat_input_ids, depth=vocab_size) output = tf.matmul(one_hot_input_ids, embedding_table) else: output = tf.nn.embedding_lookup(embedding_table, input_ids) input_shape = albert_utils_official.get_shape_list(input_ids) output = tf.reshape(output, input_shape[0:-1] + [input_shape[-1] * embedding_size]) return (output, embedding_table)
def gumbel_embedding_lookup(input_ids,
vocab_size,
embedding_size=128,
initializer_range=0.02,
word_embedding_name="word_embeddings",
use_one_hot_embeddings=False):
"""Looks up words embeddings for id tensor.
Args:
input_ids: int32 Tensor of shape [batch_size, seq_length] containing word
ids.
vocab_size: int. Size of the embedding vocabulary.
embedding_size: int. Width of the word embeddings.
initializer_range: float. Embedding initialization range.
word_embedding_name: string. Name of the embedding table.
use_one_hot_embeddings: bool. If True, use one-hot method for word
embeddings. If False, use `tf.nn.embedding_lookup()`. One hot is better
for TPUs.
Returns:
float Tensor of shape [batch_size, seq_length, embedding_size].
"""
# This function assumes that the input is of shape [batch_size, seq_length,
# num_inputs].
#
# If the input is a 2D tensor of shape [batch_size, seq_length], we
# reshape to [batch_size, seq_length, 1].
input_shape = albert_utils_official.get_shape_list(input_ids,
expected_rank=[3])
embedding_table = tf.get_variable(
name=word_embedding_name,
shape=[vocab_size, embedding_size],
initializer=create_initializer(initializer_range))
output = tf.einsum("abc,cd->abd", tf.cast(input_ids, tf.float32),
embedding_table)
return (output, embedding_table)
def dense_layer_3d(input_tensor,
num_attention_heads,
head_size,
initializer,
activation,
name=None):
"""A dense layer with 3D kernel.
Args:
input_tensor: float Tensor of shape [batch, seq_length, hidden_size].
num_attention_heads: Number of attention heads.
head_size: The size per attention head.
initializer: Kernel initializer.
activation: Actication function.
name: The name scope of this layer.
Returns:
float logits Tensor.
"""
input_shape = albert_utils_official.get_shape_list(input_tensor)
hidden_size = input_shape[2]
with tf.variable_scope(name):
w = tf.get_variable(
name="kernel",
shape=[hidden_size, num_attention_heads * head_size],
initializer=initializer)
w = tf.reshape(w, [hidden_size, num_attention_heads, head_size])
b = tf.get_variable(
name="bias",
shape=[num_attention_heads * head_size],
initializer=tf.zeros_initializer)
b = tf.reshape(b, [num_attention_heads, head_size])
ret = tf.einsum("BFH,HND->BFND", input_tensor, w)
ret += b
if activation is not None:
return activation(ret)
else:
return ret
def build_embedder(self, input_ids, token_type_ids, hidden_dropout_prob,
attention_probs_dropout_prob, **kargs):
reuse = kargs["reuse"]
input_shape = albert_utils_official.get_shape_list(
input_ids, expected_rank=[2, 3])
batch_size = input_shape[0]
seq_length = input_shape[1]
if token_type_ids is None:
token_type_ids = tf.zeros(shape=[batch_size, seq_length],
dtype=tf.int32)
if self.config.get('embedding_scope', None):
embedding_scope = self.config['embedding_scope']
other_embedding_scope = self.config.get("scope", "bert")
tf.logging.info(
"==using embedding scope of original model_config.embedding_scope: %s==",
embedding_scope)
else:
embedding_scope = self.config.get("scope", "bert")
other_embedding_scope = self.config.get("scope", "bert")
tf.logging.info(
"==using embedding scope of original model_config.scope: %s==",
embedding_scope)
with tf.variable_scope(embedding_scope, reuse=reuse):
with tf.variable_scope("embeddings"):
# Perform embedding lookup on the word ids.
(self.embedding_output_word, self.embedding_table
) = albert_modules_official.embedding_lookup(
input_ids=input_ids,
vocab_size=self.config.vocab_size,
embedding_size=self.config.embedding_size,
initializer_range=self.config.initializer_range,
word_embedding_name="word_embeddings",
use_one_hot_embeddings=self.config.use_one_hot_embeddings)
if kargs.get("perturbation", None):
self.embedding_output_word += kargs["perturbation"]
tf.logging.info(
" add word pertubation for robust learning ")
with tf.variable_scope(other_embedding_scope, reuse=reuse):
with tf.variable_scope("embeddings"):
# Add positional embeddings and token type embeddings, then layer
# normalize and perform dropout.
self.embedding_output = albert_modules_official.embedding_postprocessor(
input_tensor=self.embedding_output_word,
use_token_type=True,
token_type_ids=token_type_ids,
token_type_vocab_size=self.config.type_vocab_size,
token_type_embedding_name="token_type_embeddings",
use_position_embeddings=True,
position_embedding_name="position_embeddings",
initializer_range=self.config.initializer_range,
max_position_embeddings=self.config.
max_position_embeddings,
dropout_prob=hidden_dropout_prob,
token_type_ratio=self.config.get("token_type_ratio", 1.0))
def transformer_model(input_tensor,
attention_mask=None,
hidden_size=768,
num_hidden_layers=12,
num_hidden_groups=12,
num_attention_heads=12,
intermediate_size=3072,
inner_group_num=1,
intermediate_act_fn="gelu",
hidden_dropout_prob=0.1,
attention_probs_dropout_prob=0.1,
initializer_range=0.02,
do_return_all_layers=False):
"""Multi-headed, multi-layer Transformer from "Attention is All You Need".
This is almost an exact implementation of the original Transformer encoder.
See the original paper:
https://arxiv.org/abs/1706.03762
Also see:
https://github.com/tensorflow/tensor2tensor/blob/master/tensor2tensor/models/transformer.py
Args:
input_tensor: float Tensor of shape [batch_size, seq_length, hidden_size].
attention_mask: (optional) int32 Tensor of shape [batch_size, seq_length,
seq_length], with 1 for positions that can be attended to and 0 in
positions that should not be.
hidden_size: int. Hidden size of the Transformer.
num_hidden_layers: int. Number of layers (blocks) in the Transformer.
num_hidden_groups: int. Number of group for the hidden layers, parameters
in the same group are shared.
num_attention_heads: int. Number of attention heads in the Transformer.
intermediate_size: int. The size of the "intermediate" (a.k.a., feed
forward) layer.
inner_group_num: int, number of inner repetition of attention and ffn.
intermediate_act_fn: function. The non-linear activation function to apply
to the output of the intermediate/feed-forward layer.
hidden_dropout_prob: float. Dropout probability for the hidden layers.
attention_probs_dropout_prob: float. Dropout probability of the attention
probabilities.
initializer_range: float. Range of the initializer (stddev of truncated
normal).
do_return_all_layers: Whether to also return all layers or just the final
layer.
Returns:
float Tensor of shape [batch_size, seq_length, hidden_size], the final
hidden layer of the Transformer.
Raises:
ValueError: A Tensor shape or parameter is invalid.
"""
if hidden_size % num_attention_heads != 0:
raise ValueError(
"The hidden size (%d) is not a multiple of the number of attention "
"heads (%d)" % (hidden_size, num_attention_heads))
attention_head_size = hidden_size // num_attention_heads
input_shape = albert_utils_official.get_shape_list(input_tensor, expected_rank=3)
input_width = input_shape[2]
all_layer_outputs = []
all_attention_scores = []
if input_width != hidden_size:
prev_output = dense_layer_2d(
input_tensor, hidden_size, create_initializer(initializer_range),
None, name="embedding_hidden_mapping_in")
else:
prev_output = input_tensor
with tf.variable_scope("transformer", reuse=tf.AUTO_REUSE):
for layer_idx in range(num_hidden_layers):
group_idx = int(layer_idx / num_hidden_layers * num_hidden_groups)
with tf.variable_scope("group_%d" % group_idx):
with tf.name_scope("layer_%d" % layer_idx):
layer_output = prev_output
for inner_group_idx in range(inner_group_num):
with tf.variable_scope("inner_group_%d" % inner_group_idx):
[layer_output, attention_scores] = attention_ffn_block(
layer_output, hidden_size, attention_mask,
num_attention_heads, attention_head_size,
attention_probs_dropout_prob, intermediate_size,
intermediate_act_fn, initializer_range, hidden_dropout_prob)
prev_output = layer_output
all_layer_outputs.append(layer_output)
all_attention_scores.append(attention_scores)
if do_return_all_layers:
return all_layer_outputs, all_attention_scores
else:
return all_layer_outputs, all_attention_scores
def attention_layer(from_tensor, to_tensor, attention_mask=None, num_attention_heads=1, query_act=None, key_act=None, value_act=None, attention_probs_dropout_prob=0.0, initializer_range=0.02, batch_size=None, from_seq_length=None, to_seq_length=None): """Performs multi-headed attention from `from_tensor` to `to_tensor`. Args: from_tensor: float Tensor of shape [batch_size, from_seq_length, from_width]. to_tensor: float Tensor of shape [batch_size, to_seq_length, to_width]. attention_mask: (optional) int32 Tensor of shape [batch_size, from_seq_length, to_seq_length]. The values should be 1 or 0. The attention scores will effectively be set to -infinity for any positions in the mask that are 0, and will be unchanged for positions that are 1. num_attention_heads: int. Number of attention heads. query_act: (optional) Activation function for the query transform. key_act: (optional) Activation function for the key transform. value_act: (optional) Activation function for the value transform. attention_probs_dropout_prob: (optional) float. Dropout probability of the attention probabilities. initializer_range: float. Range of the weight initializer. batch_size: (Optional) int. If the input is 2D, this might be the batch size of the 3D version of the `from_tensor` and `to_tensor`. from_seq_length: (Optional) If the input is 2D, this might be the seq length of the 3D version of the `from_tensor`. to_seq_length: (Optional) If the input is 2D, this might be the seq length of the 3D version of the `to_tensor`. Returns: float Tensor of shape [batch_size, from_seq_length, num_attention_heads, size_per_head]. Raises: ValueError: Any of the arguments or tensor shapes are invalid. """ from_shape = albert_utils_official.get_shape_list(from_tensor, expected_rank=[2, 3]) to_shape = albert_utils_official.get_shape_list(to_tensor, expected_rank=[2, 3]) size_per_head = int(from_shape[2]/num_attention_heads) if len(from_shape) != len(to_shape): raise ValueError( "The rank of `from_tensor` must match the rank of `to_tensor`.") if len(from_shape) == 3: batch_size = from_shape[0] from_seq_length = from_shape[1] to_seq_length = to_shape[1] elif len(from_shape) == 2: if (batch_size is None or from_seq_length is None or to_seq_length is None): raise ValueError( "When passing in rank 2 tensors to attention_layer, the values " "for `batch_size`, `from_seq_length`, and `to_seq_length` " "must all be specified.") # Scalar dimensions referenced here: # B = batch size (number of sequences) # F = `from_tensor` sequence length # T = `to_tensor` sequence length # N = `num_attention_heads` # H = `size_per_head` # `query_layer` = [B, F, N, H] q = dense_layer_3d(from_tensor, num_attention_heads, size_per_head, create_initializer(initializer_range), query_act, "query") # `key_layer` = [B, T, N, H] k = dense_layer_3d(to_tensor, num_attention_heads, size_per_head, create_initializer(initializer_range), key_act, "key") # `value_layer` = [B, T, N, H] v = dense_layer_3d(to_tensor, num_attention_heads, size_per_head, create_initializer(initializer_range), value_act, "value") q = tf.transpose(q, [0, 2, 1, 3]) k = tf.transpose(k, [0, 2, 1, 3]) v = tf.transpose(v, [0, 2, 1, 3]) if attention_mask is not None: attention_mask = tf.reshape( attention_mask, [batch_size, 1, to_seq_length, 1]) # 'new_embeddings = [B, N, F, H]' new_embeddings = dot_product_attention(q, k, v, attention_mask, attention_probs_dropout_prob) return tf.transpose(new_embeddings, [0, 2, 1, 3])
def embedding_postprocessor(input_tensor,
use_token_type=False,
token_type_ids=None,
token_type_vocab_size=16,
token_type_embedding_name="token_type_embeddings",
use_position_embeddings=True,
position_embedding_name="position_embeddings",
initializer_range=0.02,
max_position_embeddings=512,
dropout_prob=0.1,
token_type_ratio=1.0):
"""Performs various post-processing on a word embedding tensor.
Args:
input_tensor: float Tensor of shape [batch_size, seq_length,
embedding_size].
use_token_type: bool. Whether to add embeddings for `token_type_ids`.
token_type_ids: (optional) int32 Tensor of shape [batch_size, seq_length].
Must be specified if `use_token_type` is True.
token_type_vocab_size: int. The vocabulary size of `token_type_ids`.
token_type_embedding_name: string. The name of the embedding table variable
for token type ids.
use_position_embeddings: bool. Whether to add position embeddings for the
position of each token in the sequence.
position_embedding_name: string. The name of the embedding table variable
for positional embeddings.
initializer_range: float. Range of the weight initialization.
max_position_embeddings: int. Maximum sequence length that might ever be
used with this model. This can be longer than the sequence length of
input_tensor, but cannot be shorter.
dropout_prob: float. Dropout probability applied to the final output tensor.
Returns:
float tensor with same shape as `input_tensor`.
Raises:
ValueError: One of the tensor shapes or input values is invalid.
"""
input_shape = albert_utils_official.get_shape_list(input_tensor, expected_rank=3)
batch_size = input_shape[0]
seq_length = input_shape[1]
width = input_shape[2]
output = input_tensor
if use_token_type:
if token_type_ids is None:
raise ValueError("`token_type_ids` must be specified if"
"`use_token_type` is True.")
token_type_table = tf.get_variable(
name=token_type_embedding_name,
shape=[token_type_vocab_size, width],
initializer=create_initializer(initializer_range))
# This vocab will be small so we always do one-hot here, since it is always
# faster for a small vocabulary.
flat_token_type_ids = tf.reshape(token_type_ids, [-1])
one_hot_ids = tf.one_hot(flat_token_type_ids, depth=token_type_vocab_size)
token_type_embeddings = tf.matmul(one_hot_ids, token_type_table)
token_type_embeddings = tf.reshape(token_type_embeddings,
[batch_size, seq_length, width])
output += token_type_ratio * token_type_embeddings
if use_position_embeddings:
assert_op = tf.assert_less_equal(seq_length, max_position_embeddings)
with tf.control_dependencies([assert_op]):
full_position_embeddings = tf.get_variable(
name=position_embedding_name,
shape=[max_position_embeddings, width],
initializer=create_initializer(initializer_range))
# Since the position embedding table is a learned variable, we create it
# using a (long) sequence length `max_position_embeddings`. The actual
# sequence length might be shorter than this, for faster training of
# tasks that do not have long sequences.
#
# So `full_position_embeddings` is effectively an embedding table
# for position [0, 1, 2, ..., max_position_embeddings-1], and the current
# sequence has positions [0, 1, 2, ... seq_length-1], so we can just
# perform a slice.
# position_embeddings = tf.slice(full_position_embeddings, [0, 0],
# [seq_length, -1])
flat_pos_ids = tf.range(seq_length, dtype=tf.int32)
one_hot_pos_ids = tf.one_hot(flat_pos_ids, depth=max_position_embeddings)
position_embeddings = tf.matmul(one_hot_pos_ids, full_position_embeddings)
num_dims = len(output.shape.as_list())
# Only the last two dimensions are relevant (`seq_length` and `width`), so
# we broadcast among the first dimensions, which is typically just
# the batch size.
position_broadcast_shape = []
for _ in range(num_dims - 2):
position_broadcast_shape.append(1)
position_broadcast_shape.extend([seq_length, width])
position_embeddings = tf.reshape(position_embeddings,
position_broadcast_shape)
output += position_embeddings
output = layer_norm_and_dropout(output, dropout_prob)
return output
